Efficient MCMC inference for remote sensing of emission sources

Abstract

A common challenge in environmental impact studies and frontier exploration is identifying the properties of some emitters from remotely obtained concentration data. For example, consider estimating the volume of some pollutant that a chemical refinery releases into the atmosphere from measurements of pollutant concentrations. Previous methods assume a known number of emitters, low ambient concentrations, or measurements from a group of stationary sensors. In contrast, we use measurements from a mobile sensor and detect source contributions that are several orders of magnitude smaller than ambient concentrations. Here, we develop and analyze a method for inferring the location, emission rate, and number of emitters from measurements taken by an aircraft. We use Reversible-jump Markov chain Monte Carlo sampling to jointly infer the posterior distribution of the number of emitters and emitter properties. Additionally, we develop performance metrics that can be efficiently computed using the sample-based representation of the posterior distribution. We investigate the expected performance of the inference algorithm with respect to certain model parameters in a series of synthetic experiments and use these performance metrics for evaluation. These experiments provide insight into subtleties of the model, including the identifiability of source configurations, the effect of various path geometries, and the effects of incorporating data from multiple flights. We also provide intuition for best-case performance when running on real-world data using a synthetic experiment. Finally, we demonstrate our ability to process and analyze real-world data for which the true source configuration is unknown.